Remedying the Inefficiencies of Clinical Trials with AI

One of the biggest uses of AI in the Pharma industry is in clinical trials. In an earlier blog post, I mentioned that twelve is the average number of years it takes for a drug to reach market, but what’s more is that half of that time and investment is spent in clinical trials.

As a refresher, once the molecule has been safely tested in animal models, there are several phases that will then be conducted in humans, beginning with Phase 0, which is largely exploratory, with a handful of patients receiving minute — most likely below therapeutic threshold — doses. Even then, many researchers may omit Phase 0 altogether as it is not a standard requirement. In Phase 1, the objective is to evaluate safety. Various doses may be tested incrementally to determine the adverse effect profile. Phase II then looks at efficacy, and may compare treatment groups (patients receiving the test molecule) against control groups (patients receiving placebo). Phase III essentially recreates phase II on a greater scale to see if the results are reproducible in a larger patient population, typically over 100 patients, depending on the disease condition studied.

Now let’s break down the timeline of a clinical trial to better understand how AI can be leveraged at each distinct stage.

Patient selection

We start with the selection of patients, and we pick individuals who seem to be most suitable for the drug in question. Patient considerations typically include age, presence of co-morbidities, previous treatment with other medications, severity of disease, and organ function. Despite establishing stringent inclusion criteria at the outset of a clinical trial, fewer than one third of phase II trials go on to phase III and an astonishing 30% of phase III trials fail because it “lacks enough patients or the right kinds of patients.” You would think that patient selection and subsequent recruitment should be a fairly smooth task, but in fact, it takes up one third of the entire clinical trial process. That alone screams inefficiency, in both time and cost.

Artificial intelligence is touted as the solution to inefficiencies of all sorts. Natural language processing, or NLP, is an artificial intelligence technique that refers to how computers make sense of written or spoken language, both the literal definition of words and the context in which they are used. Such an application, when employed in datasets in the healthcare setting — electronic health records, research literature, and prior clinical trial databases — can be used to indicate patients who best fit the design of the clinical trial, which is far quicker than the manual, laborious approach we have traditionally used whereby we rely on academic institutions, hospitals, physicians, and other healthcare professionals, or we hold out hope that patients will find out about the trial through clinical trial registries online. With NLP, we are able to canvas volumes of data to identify potential patients, and on the flip side, this same technology can assist patients, who are actively looking for clinical trials, to find a potential match. This is all to say that with the preliminary pairing of patients and clinical trials taken over by an automated process, we foresee an improvement in the patient recruitment process, and more importantly, a broader reach to the right patients.

Patient monitoring

Patient monitoring is crucial throughout each phase of the clinical trial. This is not only relevant during the study period — phases 0 to 3, — but also once the drug has received approval and entered the market, at which point, post-marketing surveillance — phase 4 — will take place. Phase 4 studies are extensive and help to elucidate long-term safety effects.

Clinical trials fail for a number of reasons. I highlighted inadequate patient recruitment as a factor in the above section, yet, another would be patient non-adherence to medications, rendering them ineligible to continue on in the trial. Non-adherence to medications may be due to forgetting, lack of understanding, or intolerable side effects. 85% of clinical trials experience some form of patient dropout and of those trials, the average dropout rate is 30%. When this occurs, trial coordinators have to recruit additional patients to maintain the integrity of the study. As you are now aware, that recruitment process will, again, cause a significant delay.

Where AI methods are expected to make an impact in this setting is through more rigorous patient monitoring and coaching, to ultimately improve support of the patients throughout the entire clinical trial. As an example, wearables have been provided to patients, which capture comprehensive, real-time data, allowing for support intervention when needed. We might immediately think of wearables like the Apple Watch and the Fitbit, but the type of wearable can be tailored to the sort of health data that is required from that particular clinical trial. For example, there are sensors that can track respiratory and cardiac measures, and others can track muscle movement or sleep. While useful to the researchers, benefits of obtaining and analyzing these data also extend to the patients themselves. There has been a considerable cultural shift in healthcare in which patients are becoming more health literate, more actively engaged in their treatment plans, and prefer to be informed about the implications and meaning behind their health data. This is no different in clinical trials and, in fact, technologies enabling increased data collection would allow patients better access to such data.

Clinical Trial Design

Now that we have discussed the supportive features of AI techniques in relation to patients, let’s take a step back and look more holistically at the design of a clinical trial. Machine learning algorithms in AI sift through data in search of patterns that we can draw conclusions from. In clinical trials, then, it is obvious that the technology can inform us of the strengths and weaknesses of previous trials and find out what factors made a trial succeed or fail.

The challenge behind leveraging AI in this manner is something that I have brought up before in earlier blog posts. Our conclusions are only as good as the data we provide. The multitude of data sources and types of data generated throughout twelve years of the average clinical trial is sure to be chaotic, unstructured, and some of it, plain insignificant. Madhusudan Shekar of Amazon Internet Services said, “Machine learning, for all the coolness it is known for, first is about data. Eighty percent of the work in machine learning is getting the data organized, structured, trustworthy. Because the machine is now going to make decisions for you.”

Pharmaceutical companies exploring the integration of AI into their clinical trials need to be aware of this; they need to consult experienced data scientists and figure out how to refine their data before feeding them into machine learning (ML) algorithms. Data sets used to train algorithms need to be filtered, accurately labelled, then categorized, and the impact of AI at each stage of the clinical trial needs to be properly assessed and measured. Put differently, substantial investments need to be made before this revolutionary technology will bear success.

Thanks for reading, as always!